Claude code agent sdk

Claude Code Agent SDK

What it is: Library for building production AI agents in TypeScript or Python with the same tools, agentic loop, and context management that power Claude Code. Same capabilities as the CLI, but programmable for headless automation, CI/CD, custom applications, and production deployments.

Packages:

• TypeScript: @anthropic-ai/claude-agent-sdk (includes native Claude Code binary)
• Python: claude-agent-sdk (requires Python 3.10+)

Installation:

npm install @anthropic-ai/claude-agent-sdk
pip install claude-agent-sdk

When to Use

Use Agent SDK when: productizing/automating agents headlessly, embedding tools into apps, running agents on schedule, building custom agents with full control, or agents need autonomous operation without interactive CLI.

Don't use when: interactive development (use CLI), one-off tasks (use CLI), or needing managed REST API without running infrastructure (use Managed Agents).

Core Entry Point: query()

import { query } from "@anthropic-ai/claude-agent-sdk";

for await (const message of query({
  prompt: "Fix the bug in auth.ts",
  options: { allowedTools: ["Read", "Edit", "Bash"] }
})) {
  console.log(message);
}

import asyncio
from claude_agent_sdk import query, ClaudeAgentOptions

async def main():
    async for message in query(
        prompt="Fix the bug in auth.py",
        options=ClaudeAgentOptions(allowed_tools=["Read", "Edit", "Bash"]),
    ):
        print(message)

asyncio.run(main())

Python also supports context manager: async with ClaudeSDKClient(options=...) as client: await client.query(...)

Key Options (Python snake_case / TypeScript camelCase)

Option	Type	Default	Purpose
allowed_tools / allowedTools	string[]	—	Pre-approve tools
disallowed_tools / disallowedTools	string[]	—	Block tools or patterns
permission_mode / permissionMode	string	"default"	Control approval: default, dontAsk, acceptEdits, bypassPermissions, plan, auto (TS only)
cwd	string	cwd	Working directory
model	string	latest Claude	Model ID
resume	string	—	Session ID to resume
max_turns / maxTurns	number	—	Max iterations
hooks	object	—	Lifecycle callbacks
agents	object	—	Subagent definitions
mcp_servers / mcpServers	object	—	MCP server config
setting_sources / settingSources	string[]	["project"]	Config sources to load

Built-in Tools

Read, Write, Edit, Bash, PowerShell, Monitor, Glob, Grep, WebSearch, WebFetch, AskUserQuestion.

Hooks: Lifecycle Callbacks

Available events (17+ total; 6 TypeScript-only):

Event	Python	TypeScript	Purpose
PreToolUse	✓	✓	Block/modify before execution
PostToolUse	✓	✓	Audit/log after completion
PostToolUseFailure	✓	✓	Handle errors
PostToolBatch	—	✓	React to batch completion
UserPromptSubmit	✓	✓	Inject context
MessageDisplay	—	✓	Transform display text
Stop	✓	✓	Save state on exit
SubagentStart	✓	✓	Track spawn
SubagentStop	✓	✓	Aggregate results
PreCompact	✓	✓	Archive before compaction
PermissionRequest	✓	✓	Custom permissions
SessionStart	—	✓	Initialize (TS only)
SessionEnd	—	✓	Cleanup (TS only)
Notification	✓	✓	Forward status
Setup	—	✓	Init tasks (TS only)
TeammateIdle	—	✓	Reassign (TS only)
TaskCompleted	—	✓	React to task (TS only)
ConfigChange	—	✓	Reload settings (TS only)
WorktreeCreate	—	✓	Track worktree (TS only)
WorktreeRemove	—	✓	Cleanup worktree (TS only)

Hooks use matcher patterns (e.g. "Write|Edit") and return hookSpecificOutput with permissionDecision, updatedInput, additionalContext, or updatedToolOutput. Note: SessionStart/SessionEnd are TypeScript-only for SDK callbacks; Python supports them only as shell command hooks in .claude/settings.json.

Subagents

Spawn specialized agents via the Agent tool. Warning: when parent uses bypassPermissions, acceptEdits, or auto mode, subagents inherit it without override—grants full system access in controlled environments only.

Sessions: Context Persistence

Capture session_id from init message, resume with resume=session_id. Sessions persist as .jsonl files locally. Use resumeSessionAt to resume at specific transcript point.

MCP Integration

Register via mcp_servers / mcpServers option or load from .claude/mcp.json / ~/.claude/mcp.json.

Permission Modes

Mode	Behavior
default	No auto-approvals; unmatched tools trigger callback
dontAsk	Deny instead of prompting
acceptEdits	Auto-approve file edits + filesystem ops
bypassPermissions	All tools run without prompts (use cautiously)
plan	Read-only tools; Claude analyzes without editing
auto	Model-classified approvals (TS only)

Evaluation order: Hooks → Deny rules → Permission mode → Allow rules → Callback.

Configuration Sources

Loaded from .claude/ and ~/.claude/ when settingSources includes those paths:

• Skills: .claude/skills/*/SKILL.md
• Commands: .claude/commands/*.md (legacy)
• Memory: CLAUDE.md or .claude/CLAUDE.md
• Settings/Hooks: .claude/settings.json
• Plugins: Programmatic via plugins option only

Authentication

Priority: api_key in options → ANTHROPIC_API_KEY env var → Third-party providers (Bedrock, AWS, Vertex AI, Azure).

Billing

Starting June 15, 2026: Agent SDK usage draws from separate monthly credit pool (distinct from interactive Claude). Credits do not roll over.

TypeScript vs Python

Concept	TypeScript	Python
Options	Passed as object	ClaudeAgentOptions
camelCase	allowedTools, mcpServers	allowed_tools, mcp_servers
Iteration	for await (const msg of query(...))	async for message in query(...)
SessionStart/End	Callback hooks	Shell command hooks only

Key Resources

• Docs: https://code.claude.com/docs/en/agent-sdk/overview
• TypeScript: https://code.claude.com/docs/en/agent-sdk/typescript
• Python: https://code.claude.com/docs/en/agent-sdk/python
• Hooks: https://code.claude.com/docs/en/agent-sdk/hooks
• Permissions: https://code.claude.com/docs/en/agent-sdk/permissions
• Subagents: https://code.claude.com/docs/en/agent-sdk/subagents
• Sessions: https://code.claude.com/docs/en/agent-sdk/sessions
• MCP: https://code.claude.com/docs/en/agent-sdk/mcp
• Plugins: https://code.claude.com/docs/en/agent-sdk/plugins
• Examples: https://github.com/anthropics/claude-agent-sdk-demos